Results of the iSEGMound Workflow

In this Master’s thesis two data preparation methods (using a plain DTM vs. a Multi-Scale Topographic Index, as described an explained in Chapter 4) and two segmentation methods (Watershed and Region Growing, as described in Chapter 2 and explained in Chapter 4) were examined, applied and compared, resulting in four workflows. The basic settings and the exact structure and process for the four workflows were tested and debugged on the Train DTM (one 1x1 km tile) and then applied and to the Train Area (five 1x1 km tile) to understand the relationship between the size of the area of investigation and the variable settings of the respective algorithms. These settings were the adjusted and the most effective workflow was chosen (based on the Train Area). This was followed by the application of the selected workflow to the five Areas of Interests: AoI 1, AoI 2, AoI 3, AoI 4 and AoI 5.

First let’s have inspect the chosen morphometric derivative, the Multi-Scale Topographic Index (later MSTPI), on the example of the Train DTM:

Multi-Scale Topographic Index of the Train DTM, Scale 1:4450.

Multi-Scale Topographic Index of the Train DTM, Scale 1:4450.

As a reminder let’s see where the burial mound groups Site ID 5 (black) and Site ID 35 (blue) are located in the Training DTM:

Multi-Scale Topographic Index of the Train DTM with bruial mound groups Site ID 5 and 35. Scale 1:4450.

Multi-Scale Topographic Index of the Train DTM with bruial mound groups Site ID 5 and 35. Scale 1:4450.

We know from Dobiat et al. 1994, that Site ID 35 was identified as two mounds. As in Chapter 4 discussed, the mounds visible in Figure 68 were possible to be identified on ground.

Results of the Training DTM

The workflows applied on the Training DTM are the following: 5a_iSEG05_WS, 5b_iSEG05_mtpi_WS, 5c_iSEG05_RG, and 5d_iSEG05_mtpi_RG.

Let’s plot the results of the Training DTM by segmentation. Left the Watershed Segmentation based on a DTM (iSEG05_WS, orange segments) and on the SAGA MTPI (iSEG05_mtpi_WS, lilac segments). Right the Region Growing Segmentation based on a DTM (iSEG05_RG, light blue) and on the SAGA MTPI (iSEG05_mtpi_RG, brown):

iSEG WS in orange and iSEG mtpi WS in lilac nect to iSEG RG in light blue and iSEG mtpi RG in brown, Scale 1:4450.iSEG WS in orange and iSEG mtpi WS in lilac nect to iSEG RG in light blue and iSEG mtpi RG in brown, Scale 1:4450.

iSEG WS in orange and iSEG mtpi WS in lilac nect to iSEG RG in light blue and iSEG mtpi RG in brown, Scale 1:4450.

The first thing that catches the eyes is that both segmentation methods were able to detect Site ID 35, using the SAGA MTPI. Thus it is already clear from this early step on, that in the case of these scarcely preserved burial mounds it is useful to work with morphometric derivatives. When comparing the two segmentation methods, it is apparent, that Watershed segmentation produces more segments than Region Growing.

Results of the Training Area

Before discussing the results of the segmentations, first let’s inspect Site ID 7 and Site ID 14.

Site ID 7 is situated relatively near to the North of Site IDs 5 and 35. The group is constituted of 9 burials, roughly in an elongated line, counted from Southwest to Northeast. When inspecting the mounds, it can be seen that, similar to Site ID 5-9, these are also very near to the forestry commuting routes. Also they already show erosion (mound Site ID 7-5 to 9), mainly in road proximity. This situation has already worsened since 2009/2010, the collection date of the LiDAR data. This burial mound group is similarly preserved such as the average height of the mounds of Site ID 5.

Site ID 14 stretches a little further away to the South and consists of altogether 18 burials. This burial mound group spreads similarly elongated as Site ID 7, although a grouping can be made out in the center region of the group. What is striking about this group is, that many of the mounds - apart from mound 8, which is cut right at the middle - have been just missed or only slightly touched by service roads. The situation of burial mound Site ID 14-8 already indicated, that it is going to be hard to detect this mound properly, because it might be will be difficult to distinguish from the road which is cutting it.

Site ID 7, consituted of 9 burial mounds and Site ID 14, constituted of 18 burial mounds on the DTM, Scale 1:1200 and 1:3100.Site ID 7, consituted of 9 burial mounds and Site ID 14, constituted of 18 burial mounds on the DTM, Scale 1:1200 and 1:3100.

Site ID 7, consituted of 9 burial mounds and Site ID 14, constituted of 18 burial mounds on the DTM, Scale 1:1200 and 1:3100.

The workflows applied on the Training Area are the following: 6a_iSEG05_WS_ta, 6b_iSEG05_mtpi_WS_ta, 6c_iSEG05_RG_ta, 6d_iSEG05_mtpi_RG_ta.

Because the Training Area is too big to really see details when plotting the whole, three plots are going to be displayed: one overview to understand the amount of segments and then the two areas containing burial mounds (Site IDs 5, 7 and 35 and Site ID 14) will be plotted next to each other to see the exact segmentation results.

Inspecting first the results of the Watershed Segmentation of the Training Area, iSEG05_WS_ta (pink segments) and iSEG05_mtpi_WS_ta (teal segments) are plotted together. It is clearly visible from the overview, that the first impression of the Training DTM is reinforced: more segments are left over by using the SAGA MTPI, which fit to min to max descriptor range as the segments complying to the burial mound mask. This means on the other hand of course more segments to check, but also more possibility to find previously not known mounds. This will be investigated in the Discussion.

Plotting iSEG WS ta and iSEG mtpi WS ta on the DTM, Scale 1:18000.

Plotting iSEG WS ta and iSEG mtpi WS ta on the DTM, Scale 1:18000.

When “zooming” in to the two areas (Figure 73) containing burial mounds, we can see the following: The Northern are (first image of Figure 74) demonstrates again the advantage of using MSTPI: the Site ID 35 is detected by the iSEG mtpi WS workflow, and also a second possible mound, which was only in the profile very slightly visible. Site ID 9 was also detected (in green), although unknowingly: only after the Whitebox MSTPI was checked against Dobiat et al. 1994, became clear that that segment might be Site ID 9. This workflow is better in detecting mounds in this area than the iSEG WS workflow, which missed Site Id 7-5,7-6,7-7 and 7-9). Looking at the Southern area (second image of Figure 74), iSEG WS workflow detected from Site ID 14 3 mounds more (14-1, 14-8 and 14-11) than the iSEG mtpi WS workflow, which detected 14-3 (but not detected by iSEG WS). Although a little less accurate in the southern area, the iSEG mtpi WS workflow is more successful.

Plotting iSEG WS ta and iSEG mtpi WS ta on the DTM, Scale 1:3000.Plotting iSEG WS ta and iSEG mtpi WS ta on the DTM, Scale 1:3000.

Plotting iSEG WS ta and iSEG mtpi WS ta on the DTM, Scale 1:3000.

Considering the Region Growing Segmentation, the overview tells us, that after filtering generally less segments are left over, which fit to min to max descriptor range as the segments complying to the burial mound mask:

Plotting iSEG RG ta and iSEG mtpi RG ta on the DTM, Scale 1:18000.

Plotting iSEG RG ta and iSEG mtpi RG ta on the DTM, Scale 1:18000.

Going into the details, iSEG_RG_ta is depicted in lime color and iSEG_mtpi_RG in grass green. It is again clear, that using the SAGA MTPI , Site ID 35 is detected, even if only the most visible one. The iSEG RG workflow does not detect all mounds from Site ID 5 (5-2 is missing and 5-5 is minimally detected), although so far all workflows detected all mounds. In the case of Site ID 7, only 7-1 (at least a part of it), 7-2, 7-3 and 7-8 was detected. The iSEG mtpi RG workflow did detect all mounds from Site ID 5, but it failed to detect Site ID 7-4, 7-7 and 7-9. Between the two workflows iSEG mtpi RG is the more successful.

Plotting iSEG RG ta and iSEG mtpi RG ta on the DTM, Scale 1:3000.Plotting iSEG RG ta and iSEG mtpi RG ta on the DTM, Scale 1:3000.

Plotting iSEG RG ta and iSEG mtpi RG ta on the DTM, Scale 1:3000.

Choosing the best fitting segmentation

We have seen, that it is clear, that SAGA MTPI as morphometric data preparation methods clearly enhances even the less well visible burial mounds and delineates the mounds more naturally. The remaining question is: how to choose between Watershed and Region growing Segmentation? Which segmentation is better? Two different considerations were investigated: the archaeological decision and the statistical decision.

From the archaeological point of view the aim is to detect as many burial mounds as possible. This can be of course broken down to the question if we want to find the exact shape of the mounds (in the case of the Training DTM and Training Area) or is the most important to detect as much as possible locations in any shape (e.g. just half or ¾ of a mound is detected) but to detect as many as possible of them. In the case of this Master’s, the archaeological choice is definately iSEG_mtpi_WS.

The only statistical measure which was found the most approximatively fitting is the Jaccard Index or Intersection over Union. This measure is used in Deep Learning as an evaluation metric to measure the accuracy of an object detection on the original data set. That is,

Demo table
Site_ID IoU_mtpi_WS IoU_mtpi_RG SUCCESS difference
Site ID 5-1 0.7133065 0.6916574 mtpi_WS 0.0216491
Site ID 5-2 0.6003428 0.3089561 mtpi_WS 0.2913867
Site ID 5-3 0.7130259 0.4737256 mtpi_WS 0.2393003
Site ID 5-4 0.5765421 0.3383026 mtpi_WS 0.2382395
Site ID 5-5 0.5115062 0.3307759 mtpi_WS 0.1807303
Site ID 5-6 0.5890535 0.4425936 mtpi_WS 0.1464599
Site ID 5-7 0.6735597 0.4919391 mtpi_WS 0.1816206
Site ID 5-8 0.5290715 0.2553324 mtpi_WS 0.2737391
Site ID 5-9 0.4605544 0.4591942 mtpi_WS 0.0013602
Site ID 7-1 NA 0.4427740 mtpi_RG NA
Site ID 7-2 0.804681 0.4973026 mtpi_WS 0.3073784
Site ID 7-3 0.6831214 0.4453112 mtpi_WS 0.2378102
Site ID 7-4 0.5397044 NA mtpi_WS NA
Site ID 7-5 0.5403348 0.3031546 mtpi_WS 0.2371802
Site ID 7-6 0.2942167 0.1334789 mtpi_WS 0.1607378
Site ID 7-7 0.432601 NA mtpi_WS NA
Site ID 7-8 0.6669532 0.3527253 mtpi_WS 0.3142279
Site ID 7-9 0.4702287 NA mtpi_WS NA
Site ID 9 0.4634704 NA mtpi_WS NA
Site ID 14-1 NA 0.3391305 mtpi_RG NA
Site ID 14-2 0.6061152 0.5231582 mtpi_WS 0.0829570
Site ID 14-3 0.5767704 NA mtpi_WS NA
Site ID 14-4 0.3778608 0.1315178 mtpi_WS 0.2463430
Site ID 14-5 0.4620573 0.2316558 mtpi_WS 0.2304015
Site ID 14-6 NA 0.1310698 mtpi_RG NA
Site ID 14-7 0.6880029 0.5178548 mtpi_WS 0.1701481
Site ID 14-8 NA NA NA NA
Site ID 14-9 0.5652943 0.2414775 mtpi_WS 0.3238168
Site ID 14-10 0.5316726 0.4640401 mtpi_WS 0.0676325
Site ID 14-11 NA 0.2050009 mtpi_RG NA
Site ID 14-12 0.5927712 0.2642519 mtpi_WS 0.3285193
Site ID 14-13 0.5579679 0.4759892 mtpi_WS 0.0819787
Site ID 14-14 0.387372 NA mtpi_WS NA
Site ID 14-15 0.3837281 0.3126799 mtpi_WS 0.0710482
Site ID 14-16 0.4654447 0.2659693 mtpi_WS 0.1994754
Site ID 14-17 0.4304659 0.3467625 mtpi_WS 0.0837034
Site ID 14-18 NA 0.2179641 mtpi_RG NA
Site ID 14-19 0.7427452 0.5368112 mtpi_WS 0.2059340
Site ID 35-1 0.5945741 0.5596460 mtpi_WS 0.0349281
Site ID 35-2 0.6019239 NA mtpi_WS NA